Closed tlortz closed 1 year ago
Hi, @tlortz ,thanks, that was fixed ~yesterday in the latest release, can you have a look here & let us know if this works better? https://www.dbdemos.ai/demo-notebooks.html?demoName=llm-dolly-chatbot
Also I tried to send a list of text to the summarizer but I don't see any performance improvement. Let me know if you have tips to increase GPU utilization!
closing this issue as we added the fix in the last release - let me know if we can improve it further!
Existing code in 02-Data-preparation notebook of llm-dolly-chatbot demo has two issues:
Recommend replacing with
This results in GPU utilization around 40% - probably low because we're using a batch size of 1, but definitely faster than using no GPU. The entire job runs in 16 minutes on the g5.4xlarge cluster created by dbdemos